Clustering and Stability
========================

simulated annealing, momentum, gradient climbing, *barrier jumping*

Noise
-----

Should gradually degrade with addition of some noise.

Don't show structure when there is none!

Don't ask user to choose K.


Benchmark Graphs
----------------

compared with kamada and fruchterman

mega-k5 (k5 with 1000 nodes in each cluster)
multi-mega-k5 (5 separate mega-k5s)
mega-k5-l2 (mega-k5 made up of five mega-k5s)
multi-mega-k5-l2
mega-k5-l3: 125k nodes, 1.3M edges, 4 levels of clustering
ladders: twists are local minima that are very difficult to escape from
ladders and blobs
multi ladders: very deep local minima -- when ladders cross
random

fruchterman worked better than kamada in all cases
barrier-jumping (VxOrd) worked best

References
----------

Pajek: http://vlado.fmf.uni-lj.si/pub/networks/Pajek
GD2001
GML File Format
GTL: Graph Template Library
http://rw2.cs.uni-sb.de/users/sander/html/gstools.html
www.sandia.gov/VxInsight
bnwylie@sandia.gov